This study explores the application of LeNet-5 convolutional neural network (CNN) on the Fashion-MNIST dataset to analyze its effectiveness in complex tasks using image recognition. The Fashion-MNIST dataset, comprising 70,000 28x28 pixel clothing images across 10 distinct categories, is constituted to offer a challenging benchmark for image recognition research. In the experiment, the data is first normalized and the labels are preprocessed using the one-hot encoding technique to adjust to the model’s input specifications. The LeNet-5 network combines several fully connected, pooling, and convolutional layers before using the Softmax layer to produce the probability distribution for each category. Through a series of experiments, it is found that LeNet-5 achieves 95% training accuracy and 90% validation accuracy on the Fashion-MNIST dataset, which not only confirms its effectiveness in modern image recognition tasks, but also demonstrates its good generalization ability. The study’s findings provide new insights for the future application of traditional CNN in the field of deep learning and provide new perspectives for future applications in more complex image recognition tasks. In addition, this study also provides the possibility for the application of LeNet-5 in scenarios with limited resources or high real-time requirements, providing valuable experience and reference for subsequent research.
